Machine Learning Quantification of Pulmonary Regurgitation Fraction from Echocardiography

Jennifer Cohen, Son Q. Duong, Naveen Arivazhagan, David M. Barris, Surkhay Bebiya, Rosalie Castaldo, Marjorie Gayanilo, Kali Hopkins, Maya Kailas, Grace Kong, Xiye Ma, Molly Marshall, Erin A. Paul, Melanie Tan, Jen Lie Yau, Girish N. Nadkarni, David Ezon

Research output: Contribution to journalArticlepeer-review

Abstract

Assessment of pulmonary regurgitation (PR) guides treatment for patients with congenital heart disease. Quantitative assessment of PR fraction (PRF) by echocardiography is limited. Cardiac MRI (cMRI) is the reference-standard for PRF quantification. We created an algorithm to predict cMRI-quantified PRF from echocardiography using machine learning (ML). We retrospectively performed echocardiographic measurements paired to cMRI within 3 months in patients with ≥ mild PR from 2009 to 2022. Model inputs were vena contracta ratio, PR index, PR pressure half-time, main and branch pulmonary artery diastolic flow reversal (BPAFR), and transannular patch repair. A gradient boosted trees ML algorithm was trained using k-fold cross-validation to predict cMRI PRF by phase contrast imaging as a continuous number and at > mild (PRF ≥ 20%) and severe (PRF ≥ 40%) thresholds. Regression performance was evaluated with mean absolute error (MAE), and at clinical thresholds with area-under-the-receiver-operating-characteristic curve (AUROC). Prediction accuracy was compared to historical clinician accuracy. We externally validated prior reported studies for comparison. We included 243 subjects (median age 21 years, 58% repaired tetralogy of Fallot). The regression MAE = 7.0%. For prediction of > mild PR, AUROC = 0.96, but BPAFR alone outperformed the ML model (sensitivity 94%, specificity 97%). The ML model detection of severe PR had AUROC = 0.86, but in the subgroup with BPAFR, performance dropped (AUROC = 0.73). Accuracy between clinicians and the ML model was similar (70% vs. 69%). There was decrement in performance of prior reported algorithms on external validation in our dataset. A novel ML model for echocardiographic quantification of PRF outperforms prior studies and has comparable overall accuracy to clinicians. BPAFR is an excellent marker for > mild PRF, and has moderate capacity to detect severe PR, but more work is required to distinguish moderate from severe PR. Poor external validation of prior works highlights reproducibility challenges.

Original languageEnglish
JournalPediatric Cardiology
DOIs
StateAccepted/In press - 2024

Keywords

  • Cardiac magnetic resonance imaging
  • Congenital heart disease
  • Echocardiography
  • Machine learning
  • Model validation
  • Pulmonary regurgitation
  • Tetralogy of Fallot
  • XGBoost

Fingerprint

Dive into the research topics of 'Machine Learning Quantification of Pulmonary Regurgitation Fraction from Echocardiography'. Together they form a unique fingerprint.

Cite this